axe MCP Server Version 1.5.0 Release Notes
Release Date
August 27, 2026
Overview
This v1.5.0 release makes Intelligent Guided Tests (IGTs) return dramatically smaller responses, cutting the token cost your coding agent pays to read them. It also brings IGTs into the analyze tool — one call now runs the axe scan and an IGT against the same page — extends the screenshots added in v1.4.0 with an option to save them to disk, and keeps long OAuth sessions alive without a restart.
To get this release:
- Docker — re-pull the image:
docker pull dequesystems/axe-mcp-server:latest - npm — no separate install step is required;
npxruns the latest published version. See Choosing a Distribution.
New Features
Intelligent Guided Tests in the analyze Tool
The analyze tool now accepts an optional igtTools array that runs Automated Intelligent Guided Tests against the same page, in the same call, right after the axe scan:
{
"url": "http://localhost:3000",
"igtTools": ["keyboard"]
}Scanning a page and running the Keyboard IGT against it previously took two separate tool calls — two browser launches, two navigations, and two rounds of before steps to reach the same state. Now everything that prepares the page runs once and carries into both. The Keyboard IGT is the supported value today.
A call that omits igtTools is unchanged, so existing prompts and agent instructions keep working untouched. IGTs consume AI credits; the axe scan does not.
See Intelligent Guided Tests for the response shape and full usage.
Screenshots Can Now Be Saved to Disk
The screenshot parameter introduced in v1.4.0 returned the image inline only. It can now also write the capture to a file:
{
"url": "http://localhost:3000",
"screenshot": { "saveTo": "/Users/me/Desktop/home.png" }
}This makes the capture reachable in MCP clients that don't render inline images. Pairing a save with inline: false also skips the inline image block entirely, so you get the screenshot without paying the image-input tokens it would cost your agent on its next turn.
See Saving screenshots to disk for the full options and the Docker volume requirement.
Uninterrupted OAuth Sessions
OAuth access tokens are short-lived, so an agent session that ran for hours would outlive its token and start failing — recoverable only by fetching a new token and restarting the server.
The new @deque/axe-auth run command removes that cliff. Point your MCP client at it as the server command, and it launches and supervises the server while keeping its access token fresh for the whole session. Your refresh token never leaves your machine, and it works for both distributions on macOS, Windows, and Linux.
See Keeping a long session alive for the configuration.
Changes
Smaller IGT Responses
IGT responses now return each processed element trimmed to the fields an agent actually needs to identify and describe it. On pages with many interactive elements these payloads had grown large enough to crowd the model's context window. Reported issue and element counts are unchanged.
This can reduce your coding agent's token consumption. Every IGT response is read by your agent's underlying model — Claude, GitHub Copilot, Cursor, and so on — and you pay that provider for the tokens it spends processing it. A smaller response means fewer tokens on every IGT run.
This is separate from AI credits, Deque's own currency for AI-powered features. Credit consumption for an IGT run is unchanged.
The Standalone igt Tool Is Deprecated
Now that analyze runs IGTs directly, the standalone igt tool is deprecated. It remains fully functional and returns the same results as before, but it will be removed in a future release.
No action is required today. If you maintain custom agent instructions that name the igt tool, update them to call analyze with igtTools instead.
Bug Fixes
- Intelligent Guided Tests no longer stall silently. When the browser extension stopped responding, a run could hang until your MCP client gave up — with no error and nothing in the log to explain it. It now fails promptly with a message you can act on.
- API key validation is more reliable on restrictive networks. Validation took a different network path than the rest of the server, so it could fail on corporate networks where every other request to Deque succeeded. It now uses the same path as everything else. If you still see connectivity problems reported as API key errors, see Troubleshooting.
